Osegi
AI·2 min read

Tiny 2.1: an AI language model that runs on a 486 from a floppy

A Reddit user built Tiny 2.1, a small language model that boots on MS-DOS, runs on a 486DX and fits on a 1.44 MB floppy. Another AI wrote the code.

A vintage beige PC running a neon text interface from a 1.44 MB floppy disk, with a tiny AI chatbot on screen.
Image: xataka.com

Running a chatbot on a computer from the early 1990s sounds like a joke. Tiny 2.1 is the real thing: a working language model that runs on MS-DOS, on a Pentium or a 486DX, and fits on a 1.44 MB floppy disk.

The project was shared on the ClaudeAI subreddit by a user called wollymonjak, and the full thing is on GitHub. According to the author, the code was written by Claude Fable 5.2 and the model was trained by Claude Opus 5.5. The program itself is in Turbo C 2.0, a compiler from the late 1980s.

How big is Tiny 2.1?

Small, obviously. It has 367,000 parameters spread over 4 layers that are 96 units wide. It works letter by letter, not word by word. The weights are stored in 8 bits, so the whole model weighs 380 KB, and it needs about 540 KB of free conventional memory.

It uses the same basic architecture as ChatGPT or Claude, just at a tiny scale. It runs in a full-screen neon interface with a live activity monitor and a robotic voice.

What can it do?

Not much, and that is fine. It is a question-and-answer model trained on roughly 3,500 facts, each phrased in many different ways. Ask about anything outside that data and it makes up an answer. The people who tried it in an online MS-DOS emulator found that long or complex questions get nonsense back. The charm is that a transformer works at all in a space this small.

How to run it

Go to the project's GitHub page, click the green Code button and download the ZIP. Then upload it to any online MS-DOS emulator and use the commands listed by the author. It can be recompiled with TCC -ml TINY.C. Write to it in English, which lowers the odds of a random answer.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:tiny 2.1ms-dos486language modelretro computingclaude

More from this day

See all →

More in: AI

See all →